A Robust Optimal Mean Cosine Angle 2DPCA for Image Feature Extraction

Pengfei Bi,Yiyan Deng,Xue Du
DOI: https://doi.org/10.1007/s00521-022-07572-z
2022-01-01
Neural Computing and Applications
Abstract:Currently, angle two-dimensional principal component analysis (Angle 2DPCA) effectively enhances the robustness of traditional 2DPCA by using a measurement model based on F-norm in the relationship between reconstruction error and variance, and successfully applied in image feature extraction. However, the vital disadvantages of Angle 2DPCA are low computational efficiency, imperfect robustness and poor interpretability. To overcome these problems, we propose robust optimal mean cosine angle 2DPCA (ROMCA-2DPCA), a novel formulation of Angle 2DPCA. ROMCA-2DPCA employs R-1-norm to maximize the ratio between the variance and input data and considers the optimal mean to complete data centralization, thereby not only reasonably reducing the computational complexity of the model but also effectively improving its robustness. Moreover, ROMCA-2DPCA strictly proves its convergence in theory, thereby greatly enhancing the interpretability of the model. Accordingly, we have developed a fast nongreedy iterative algorithm for solving the corresponding criterion function. Finally, the experimental results on some image databases show that our model is superior.
What problem does this paper attempt to address?